Changed MPI->multivariate for computing each SPI to match the bivariate signature. Also fixed the pdist signature (it was adjacency)

Author: olivercliff

demos/demo.py

@@ -7,14 +7,31 @@
 
 import seaborn as sns
 
-calc = Calculator(dataset=load_dataset('forex'))
+# Load one of our stored datasets
+dataset = load_dataset('forex')
+
+# visualize the dataset as a heat map (also called a temporal raster plot or carpet plot)
+plt.pcolormesh(dataset.to_numpy(squeeze=True),cmap='coolwarm',vmin=-2,vmax=2)
+plt.show()
+
+# Instantiate the calculator (inputting the dataset)
+calc = Calculator(dataset=dataset)
+
+# Compute all SPIs (this may take a while)
 calc.compute()
+
+# Now, we can access all of the matrices by calling calc.table.
+# This property will be an Nx(SN) pandas dataframe, where N is the number of processes in the dataset and S is the number of SPIs
+print(calc.table)
+
+# We can use this to compute the correlation between all of the methods on this dataset...
 corrmat = calc.table.stack().corr(method='spearman').abs()
 
+# ...and plot this correlation matrix
 sns.set(font_scale=0.5)
 g = sns.clustermap( corrmat.fillna(0), mask=corrmat.isna(),
                     center=0.0,
                     cmap='RdYlBu_r',
                     xticklabels=1, yticklabels=1 )
 plt.setp(g.ax_heatmap.xaxis.get_majorticklabels(), rotation=45, ha='right')
-plt.show()
+plt.show()

pyspi/base.py

@@ -89,7 +89,7 @@ def bivariate(self,data,i=None,j=None):
         raise NotImplementedError("Method not yet overloaded.")
 
     @parse_multivariate
-    def mpi(self,data):
+    def multivariate(self,data):
         """ Compute the dependency statistics for the entire multivariate dataset
         """
         A = np.empty((data.n_processes,data.n_processes))
@@ -135,8 +135,8 @@ def ispositive(self):
         return False
 
     @parse_multivariate
-    def mpi(self,data):
-        A = super(undirected,self).mpi(data)
+    def multivariate(self,data):
+        A = super(undirected,self).multivariate(data)
 
         li = np.tril_indices(data.n_processes,-1)
         A[li] = A.T[li]

pyspi/calculator.py

@@ -16,7 +16,7 @@
 
 """ TODO: use the MPI class for each entry in the table
 """
-class MPI():
+class multivariate():
     def __init__(self, procnames, S=None):
         if S is None:
             S = np.full((len(procnames),len(procnames)),np.nan)
@@ -158,14 +158,14 @@ def compute(self,replication=None):
             replication = 0
 
         pbar = tqdm(self.spis.keys())
-        for m in pbar:
-            pbar.set_description(f'Processing [{self._name}: {m}]')
+        for spi in pbar:
+            pbar.set_description(f'Processing [{self._name}: {spi}]')
             start_time = time.time()
             try:
-                self._table[m] = self._spis[m].mpi(self.dataset)
+                self._table[spi] = self._spis[spi].multivariate(self.dataset)
             except Exception as err:
-                warnings.warn(f'Caught {type(err)} for SPI "{self._statnames[m]}": {err}')
-                self._table[m] = np.NaN
+                warnings.warn(f'Caught {type(err)} for SPI "{spi}": {err}')
+                self._table[spi] = np.NaN
         pbar.close()
 
     def rmmin(self):

pyspi/statistics/basic.py

@@ -42,7 +42,7 @@ def _from_cache(self,data):
         return mycov
 
     @parse_multivariate
-    def mpi(self,data):
+    def multivariate(self,data):
         mycov = self._from_cache(data)
         matrix = getattr(mycov,self._kind+'_')
         np.fill_diagonal(matrix,np.nan)

pyspi/statistics/causal.py

@@ -120,7 +120,7 @@ def _from_cache(self,data):
         return ccmf
 
     @parse_multivariate
-    def mpi(self,data):
+    def multivariate(self,data):
         ccmf = self._from_cache(data)
 
         if self._statistic == 'mean':

pyspi/statistics/distance.py

@@ -20,7 +20,7 @@ def __init__(self,metric='euclidean',**kwargs):
         self.name += f'_{metric}'
 
     @parse_multivariate
-    def adjacency(self,data):
+    def multivariate(self,data):
         return pairwise_distances(data.to_numpy(squeeze=True),metric=self._metric)
 
 """ TODO: include optional kernels in each method

pyspi/statistics/spectral.py

@@ -77,7 +77,7 @@ def _get_cache(self,data):
         return res, freq
 
     @parse_multivariate
-    def mpi(self, data):
+    def multivariate(self, data):
         adj_freq, freq = self._get_cache(data)
         freq_id = np.where((freq >= self._fmin) * (freq <= self._fmax))[0]
 
@@ -299,7 +299,7 @@ def _get_statistic(self,C):
 #         self.name = self.name + paramstr
 
 #     @parse_multivariate
-#     def mpi(self,data):        
+#     def multivariate(self,data):        
 #         # This should be changed to conditioning on all, rather than averaging all conditionals
 #         if not hasattr(data,'pcoh'):
 #             z = np.squeeze(data.to_numpy())
@@ -372,7 +372,7 @@ def _get_cache(self,data):
         return F, freq
 
     @parse_multivariate
-    def mpi(self,data):
+    def multivariate(self,data):
         try:
             F, freq = self._get_cache(data)
             freq_id = np.where((freq >= self._fmin) * (freq <= self._fmax))[0]
@@ -399,7 +399,7 @@ def __init__(self,orth=False,log=False,absolute=False):
             self.name += '_abs'
 
     @parse_multivariate
-    def mpi(self, data):
+    def multivariate(self, data):
         z = np.moveaxis(data.to_numpy(),2,0)
         adj = np.squeeze(mnec.envelope_correlation(z,orthogonalize=self._orth,log=self._log,absolute=self._absolute))
         np.fill_diagonal(adj,np.nan)

pyspi/statistics/wavelet.py

@@ -59,7 +59,7 @@ def _get_cache(self,data):
         return conn, freq_id
 
     @parse_multivariate
-    def mpi(self, data):
+    def multivariate(self, data):
         adj_freq, freq_id = self._get_cache(data)
         try:
             adj = self._statfn(adj_freq[...,freq_id,:], axis=(2,3))
@@ -189,7 +189,7 @@ def _get_cache(self,data):
         return psi, freq_id
 
     @parse_multivariate
-    def mpi(self, data):
+    def multivariate(self, data):
         adj_freq, freq_id = self._get_cache(data)
         adj = self._statfn(np.real(adj_freq[...,freq_id]), axis=(2,3))
 

test/test_statistics.py

@@ -53,7 +53,7 @@ def test_yaml():
     assert calc.n_statistics == len(calc._statistics), (
                 'Property not equal to number of statistics')
 
-def test_mpi():
+def test_multivariate():
     # Load in all base statistics from the YAML file
 
     data = get_data()
@@ -81,14 +81,14 @@ def test_mpi():
         if any([m.name == e for e in excuse_stochastic]):
             continue
 
-        m.mpi(get_more_data())
+        m.multivariate(get_more_data())
 
-        scratch_adj = m.mpi(data.to_numpy())
-        adj = m.mpi(data)
+        scratch_adj = m.multivariate(data.to_numpy())
+        adj = m.multivariate(data)
         assert np.allclose(adj,scratch_adj,rtol=1e-1,atol=1e-2,equal_nan=True), (
                     f'{m.name} ({m.humanname}) mpi output changed between cached and strach computations: {adj} != {scratch_adj}')
 
-        recomp_adj = m.mpi(data)
+        recomp_adj = m.multivariate(data)
         assert np.allclose(adj,recomp_adj,rtol=1e-1,atol=1e-2,equal_nan=True), (
                     f'{m.name} ({m.humanname}) mpi output changed when recomputing.')
 
@@ -111,7 +111,7 @@ def test_mpi():
                     assert t_s == pytest.approx(new_t_s,rel=1e-1,abs=1e-2), (
                         f'{m.name} ({m.humanname}) Bivariate output from cache mismatch results from scratch for computation ({j},{i}): {t_s} != {new_t_s}')
                 except NotImplementedError:
-                    a = m.mpi(p[[i,j]])
+                    a = m.multivariate(p[[i,j]])
                     s_t, t_s = a[0,1], a[1,0]
 
                 if not math.isfinite(s_t):
@@ -206,7 +206,7 @@ def test_group():
     test_yaml()
     test_load()
     test_group()
-    test_mpi()
+    test_multivariate()
 
     # This was a bit tricky to implement so just ensuring it passes a test from the creator's website
     test_ccm()